Agentoptosis: Hierarchical Telehomeostasis in Persistent Meta-Agents

Programmed death, reframed as a regulatory operation: when a collective agent should eliminate one of its own parts, how it decides, and why that decision is hard to fake.

Every living cell has machinery dedicated to killing itself. That's strange if you think of the cell as the relevant unit — why would evolution preserve a suicide kit? The standard answer is that the cell isn't the relevant unit. The organism is. A cell that destroys itself at the right moment can benefit the larger pattern it belongs to more than a cell that clings to existence. This paper takes that intuition and asks: can we state it precisely, across all substrates — cells, ant colonies, software systems — without relying on biology-specific vocabulary?

The answer is yes, and the key move is to define a deletion value Γ: the counterfactual change in the collective's persistence if you intervene to remove a constituent versus keeping it. Agentoptosis (the paper's coinage, deliberately distinct from "apoptosis" to avoid confusion with the existing engineering literature) is what happens when a regulated policy fires because Γ is positive — the collective persists better without this part. The paper is careful to separate three things that are easy to conflate: the true Γ (a fact about the world, uncomputable), the collective's estimate of Γ (what sensors and signals actually track), and the long-run value of the policy that fires on that estimate (what selection actually sees). The criterion for a good agentoptotic policy lives at the third level, not the first.

Two quantities govern the decision. The first is χ, the unique information a constituent carries about the beneficiary pattern — how much harder the collective becomes to describe once this part is gone. Small χ means the part is informationally redundant; destroying it destroys an instance, not the pattern. The second is functional replacement cost, which χ cannot capture: a part may be re-specifiable in principle but unreachable in the time that matters. Together these set V, the cost of a false positive (killing something useful). Against V sits Λ, the expected damage from retaining a harmful constituent. The deletion threshold is then θ = V/(V+Λ), a standard Bayes decision rule whose arguments now have a precise persistence-theoretic reading. This single ratio retrodicts the known gradient of apoptotic thresholds across tissues — gut epithelium turns over in days (low V, high Λ), adult neurons almost never die (high V, low Λ) — without fitting each tissue separately.

The paper also asks how a constituent should be eliminated, not just whether. When a cell dies, what happens to its contents matters: are they packaged for recycling, released as an alarm signal, converted into a weapon, or just dumped? The paper calls these "discard channels" and argues that channel selection is itself under regulation, driven by the cost of uncontrolled release. Dense tissue selects for contained, phagocyte-legible death (classical apoptosis); dilute bacterial populations just lyse; termite soldiers rupture and coat attackers with toxic paste because the debris is the weapon. The channel is not an afterthought — in some cases (skin barrier formation, xylem water conduits, neutrophil chromatin traps) the dead constituent is the product, so containment would destroy the function entirely.

Finally, the paper confronts the obvious game-theoretic problem: any death switch creates a mutant opportunity. A constituent that disables its own elimination machinery pays no cost and free-rides on everyone else's compliance. Agentoptosis is therefore only stable where within-group conflict is separately suppressed — high relatedness, reproductive bottlenecks, policing. Cancer is reframed here not just as a threshold-detection failure but as somatic selection for escape variants, which is a different and harder problem. A closing speculative section applies the same logic to human social isolation: perceived burdensomeness reads as an inflated estimate of one's deletion value, and isolation removes the reciprocal social evidence that would correct the estimate — a miscalibrated constituent-evaluation mechanism, not an adaptive signal.